Why AI researchers stopped publishing their extinction claims

Geoffrey Irving, formerly at OpenAI and Google DeepMind, recently told journalists that the probability of human extinction from superintelligence is about 50 percent — a coin flip. Alongside him, a dozen other researchers from the same three labs released similar assessments through Palisade Research, a nonprofit collecting and amplifying these warnings. But the extinction probability exists nowhere else: not in a paper Irving published, not in methodology Palisade released, not in any document outside the framing of the interview itself.

This is not new. It is the 2015-2017 AI safety discourse arms race, running again.

From roughly 2015 onward, as Nick Bostrom's Superintelligence gained citation weight in tech, researchers from competing labs began issuing public warnings about existential AI risk. Each statement carried the shape of expertise: named researchers, institutional affiliation, grave tone. None carried the thing expertise requires — publishable methodology, quantified assumptions, peer review, evidence that could be wrong in specific ways. What they carried instead was recruitment power. When a researcher's warning lands in major media, that researcher's lab gains gravity. Funding follows. The warning becomes a signal that the lab takes the problem seriously, which becomes a signal that working there matters. The public claim substitutes for the scholarly one, and the distinction collapses.

What nobody asked then. What the coverage avoids now: Do OpenAI, Google, and Anthropic actually endorse these 50 percent estimates? What methodology supports the number? Why did these researchers stop publishing and start performing? The story treats the warning as a fact about superintelligence when it is a fact about institutional incentives. About researchers from three labs competing for the same moral authority by making the same undefended claim simultaneously.

When you see expertise expressed entirely through media channels rather than through the friction of peer review, you are not watching knowledge production. You are watching institutional recruitment disguised as crisis communication. The pattern has run before. It ends not in consensus about risk but in external skepticism about whether any of this was ever grounded in evidence at all. That skepticism is warranted. Because the researchers made it warranted the moment they chose journalists over journals.